fish-speech
fish-speech copied to clipboard
Brand new TTS solution
用的docker部署, ``` docker run -it --gpus all -v /ai/fish-speech/checkpoints:/exp/checkpoints -p 7899:7860 -p 8022:8000 lengyue233/fish-speech bas python -m tools.api --listen 0.0.0.0:8000 --llama-checkpoint-path "checkpoints/fish-speech-1.2" --decoder-checkpoint-path "checkpoints/fish-speech-1.2/firefly-gan-vq-fsq-4x1024-42hz-generator.pth" --decoder-config-name firefly_gan_vq ``` ![image](https://github.com/fishaudio/fish-speech/assets/40717349/8dfa6c85-dcaf-4286-8f19-bd494cb9cd54) 请求案例: ```...
**Describe the bug** Can't open inference server. **To Reproduce** 1. Run install_env.bat with USE_MIRROR=false and INSTALL_TYPE=stable 2. Change API_FLAGS.txt and enable "--infer", then Run start.bat 3. Go to the inference...
I don't need some of the generated audios from https://fish.audio. However, I haven't found a button to delete. _Originally posted by @limingchina in https://github.com/fishaudio/fish-speech/discussions/352_
**前提** 我看了介绍,目前的体系是不基于音素的,直接用文字会导致一些非 常用字 于样本上无法覆盖 **方案** 我的想法是,目前的大部分方言都有拼音方案,而且想要支持方言的人大部分也比较了解方言拼音。 可以增加一个叫"方言"的语音选项,样本标注文本就填拼音就好,每个字的拼音直接用空格分隔就好,和英文单词的分割一样,非常好处理。这样就可以支持所有的汉语方言。个人去改代码添加语言支持也是很麻烦的事,可能会漏掉很多需要修改的地方。 考虑到常用汉字都好几千,正常文本用到的至少7000+,还有多音字的处理等等麻烦事;而实际上绝大部分汉语方言,带声调的音节数量,总共最多也就2000,甚至北方方言一般1000以下。这是用拼音相对于用汉字的好处。 如此,需要的语音样本数量将非常少,就能覆盖全部音节。 此外,方言训练本身最大的问题就是获取语音样本困难,除了粤语,其它语言很难获取到足够的语音样本数据。如果所需要的样本少,甚至可以自己录制就够了。现在开源社区中,也已经有了一些方言输入法,个人想办法把汉字转成方言拼音也是比较简单的。 以上
如图所示,跟着文档一步步来的,是不是哪个地方没做对? ![image](https://github.com/fishaudio/fish-speech/assets/86335228/d4b5f280-2393-47ea-8480-a4831554850d)
**Is this PR adding new feature or fix a BUG?** Add feature. **Is this pull request related to any issue? If yes, please link the issue.** #358
Feel free to ask any kind of questions in the issues page, but please use English since other users may find your questions valuable. **Describe the bug** A clear and...
https://github.com/sh-lee-prml/HierSpeechpp Any comparison to this approach?
不使用lora训练的ckpt,load后出现Negative code found: {codes}, 生成codes出现0,但是训练loss正常 如果使用lora训练后merge的模型,一切正常
I want to finetune it with around 30mins of english singaporean accent data, but when the models was merge with lora after trained, i test the model with "hello my...